AI agents for drug discovery: where they stop
An honest page about a boundary. Agentic systems are being pointed at discovery from several directions at once, and the interesting question is not what they can attempt but where the loop closes. The system described on this site does not enter discovery at all.
Start with the boundary, because it decides whether the rest is worth your time. The product behind this site is an operations layer: it reads a person's mail and meetings, files them, keeps the record current, and tracks commitments. It does not do target identification, molecule design, screening, structure prediction or data analysis, and no amount of configuration turns it into something that does. If you came looking for a discovery platform, this is not one.
What an agent is, stripped of the marketing
An agent is a model that can call tools, observe what came back, and decide what to do next, repeatedly, without a person in each loop. That is the whole idea. Everything interesting follows from two questions: which tools it can reach, and how a claim it produces gets checked before anyone acts on it.
In software, both questions have easy answers. The tools are code execution and file access, and the check is that the program either runs or does not. Discovery has neither property. The tools that matter most are physical, and the check is an experiment that takes weeks and costs real money.
Where the loop stops closing
Follow a discovery question end to end and you can see the point where autonomy runs out.
| Step | Can a loop close on its own? |
|---|---|
| Gather and summarize what is already published | Substantially, with the usual caveat that a confident summary is not a verified one |
| Propose candidates, designs or hypotheses | Yes, and abundantly. Generating options was never the scarce thing |
| Run computational evaluation over those options | Yes, within the limits of whatever the model or simulation is actually predictive of |
| Test in a real biological system | No. This is a physical process with cost, scheduling, materials and people attached |
| Decide what the result means for the program | No, and mostly it should not. This is where a company commits money |
The last two rows are why "agents for drug discovery" reads differently to a computational chemist than it does to someone building software. The scarce resource in a discovery program is not ideas, and it is not compute. It is validated evidence, produced slowly, in an assay somebody had to develop, in a system that may or may not represent the disease. An agent that generates a hundred plausible directions has not moved the program forward, it has moved the bottleneck one step to the right, into the queue for the lab.
What this page is not. It is not a survey of the field, and it deliberately names no products, benchmarks or results. If you are here from the academic side looking for applications and case studies, the primary literature is the right source and a marketing page is not. What is offered here is one boundary argument, stated so that nobody arrives at the rest of this site with the wrong expectation.
The part of discovery that is not discovery
There is a second thing in the way, and it is the thing this site is actually about. A discovery program at a small company is also an administrative object. It has a collaborator who owes you a construct, a contract lab running a screen against a statement of work, a materials transfer agreement with a notice period, a scientific advisory board whose recommendation from the last meeting was never minuted, and a decision from March about which chemical series to carry that nobody can now reconstruct.
None of that is science and all of it slows a program down. It is also the part where an agent's loop does close, because the tools involved are email, calendars, files and records, and the check is cheap: a human reads a drafted reply before it goes anywhere. That argument in full is the operations half of drug development, and its functional version for a research leader is what a biotech CSO actually needs.
If you are evaluating both
Keep the two purchases separate. A discovery platform is bought on whether its predictions hold up in your hands, on your targets, and that evaluation is scientific. A coordination layer is bought on whether your records are current on a Tuesday without anybody maintaining them, and that evaluation takes two weeks of ordinary work to judge. Buying one hoping it will quietly do the other is how both disappointments happen.
For the same boundary drawn on the commercial side of a large organization, AI agents for pharma takes the pharmacovigilance question directly. For what a self-maintaining record actually looks like once it exists, the concept page is an AI company brain, and the function pages sit under AI for life sciences.
Where the loop does close
Nothing on targets, molecules or assay data. A real app, walked through in the demo, holding the collaborator who owes you a construct, the screen running to a statement of work, the notice period in the transfer agreement and the March series decision nobody can now reconstruct. Watch it, and if that is not what slows your program down, better to know now.
See the demo